Power grid planning construction operation integrated modeling method and system based on semantic system

Abstract

The application provides a power grid planning construction operation integration modeling method and system based on a semantic system, relates to the technical field of power grid planning, comprises semantic labeling of power grid planning historical data, construction of a semantic vector space, generation of a multi-dimensional semantic feature library, and feature extraction, fusion and decision output by using a multi-layer neural network. The method receives real-time power grid planning demand data, outputs an initial scheme containing load distribution prediction results, network structure configuration scheme and investment scale suggestion, and then iteratively optimizes by using a reinforcement learning algorithm, and finally outputs an optimal power grid planning scheme. The application can effectively fuse multi-source heterogeneous power grid data, improve the accuracy and efficiency of power grid planning, reduce investment cost, and improve the reliability, economy and environmental protection of the power grid.

Description

TECHNICAL FIELD

[0001] The present application relates to power grid planning technology, in particular to a power grid planning construction operation integration modeling method and system based on a semantic system. BACKGROUND

[0002] Traditional power grid planning construction operation methods mainly rely on expert experience and manual analysis, using static planning methods, which are difficult to adapt to the challenges brought by the continuous expansion of the power grid scale, rapid growth of load and large-scale access of new energy. The three stages of planning, construction and operation are relatively independent, lacking effective information sharing and collaboration mechanisms, leading to a disconnect between the planning scheme and actual construction and operation, making it difficult to maximize the whole life cycle benefits. In addition, the traditional planning method does not make full use of the massive historical data, and it is difficult to fully realize the value of data, leading to the lack of scientificity and foresight in the formulation of the planning scheme.

[0003] Traditional power grid planning methods are difficult to effectively handle complex and variable power grid operating environments, and do not fully consider uncertain factors such as load forecasting and new energy output fluctuations, resulting in poor robustness and adaptability of the planning scheme. The existing planning method lacks overall consideration of the whole process of planning, construction and operation, and the information island phenomenon between stages is serious, making it difficult to achieve optimal allocation and collaborative operation of resources, affecting the overall operation efficiency of the power grid. The traditional planning method does not make full use of the massive data, and it is difficult to effectively extract the valuable information contained in the data, leading to the lack of data support for the formulation of the planning scheme, and it is difficult to achieve precise planning and scientific decision-making. SUMMARY

[0004] The embodiment of the present application provides a power grid planning construction operation integration modeling method and system based on a semantic system, which can solve the problems in the prior art.

[0005] The first aspect of the embodiment of the present application is,

[0006] A power grid planning construction operation integration modeling method based on a semantic system is provided, comprising:

[0007] Obtain power grid planning historical data, the power grid planning historical data including load forecasting data, power grid topology data and equipment parameter data; perform semantic annotation on the power grid planning historical data, establish a power grid planning field glossary; construct a semantic vector space based on the power grid planning field glossary, map the power grid planning historical data into the semantic vector space, and generate power grid planning basic semantic features; according to the power grid planning business rules, the power grid planning basic semantic features are multi-dimensionally decomposed to form a multi-dimension semantic feature library including time dimension features, space dimension features and business dimension features;

[0008] A multi-layer neural network structure is constructed, which includes a feature extraction layer, a feature fusion layer and a decision output layer; the feature data in the multi-dimensional semantic feature library is input into the feature extraction layer, and local features are extracted through convolution operation; in the feature fusion layer, an attention mechanism is used to adaptively allocate weights and fuse features of different dimensions; in the decision output layer, a multi-task learning module is set up, and model training is performed for power grid load forecasting, grid structure optimization and investment benefit evaluation respectively;

[0009] Real-time power grid planning demand data is received, and the real-time power grid planning demand data is input into the trained deep learning model; an initial power grid planning scheme is output based on the deep learning model, which includes load distribution prediction results, grid structure configuration scheme and investment scale suggestion; a reinforcement learning algorithm is used to iteratively optimize the initial power grid planning scheme, an optimization objective function is constructed according to power grid reliability index, economy index and environmental protection index, and an optimal power grid planning scheme is output.

[0010] Power grid planning historical data is obtained, which includes load prediction data, power grid topology data and equipment parameter data; the power grid planning historical data is semantically annotated, and a power grid planning field glossary is established, including:

[0011] Power grid planning historical data is obtained through a data acquisition interface, which includes load prediction data, power grid topology data and equipment parameter data; a data quality scoring index system is established for the power grid planning historical data, which includes data integrity index, data accuracy index and data timeliness index; the quality of the power grid planning historical data is evaluated based on the data quality scoring index system, and a quality evaluation report is generated;

[0012] According to the evaluation results in the quality evaluation report, the power grid planning historical data is processed in stages, and the power grid planning historical data is divided into valid data, repairable data and invalid data; interpolation methods based on time series correlation are used to repair missing values in the repairable data, and statistical characteristics and expert rules are used to identify and correct abnormal values in the repairable data, generating repaired data; the valid data and the repaired data are combined to form preprocessed data;

[0013] A unified data model is designed, which includes data identification field, attribute description field and association relationship field; the preprocessed data is mapped to the unified data model through ETL conversion rules to form standardized data; version identification is established for the standardized data to record data change information and generate data traceability chain;

[0014] extracting a power grid planning basic term from a normative document, taking the basic term as an initial term; analyzing the standardized data by using a term frequency-inverse document frequency algorithm, and extracting a feature vocabulary; submitting the feature vocabulary to a field expert for auditing, and merging the feature vocabulary that passes the auditing with the initial term to form an initial vocabulary table;

[0015] constructing a multi-level semantic annotation system including a part-of-speech annotation layer, a concept category annotation layer, and a relationship annotation layer; performing part-of-speech annotation on the initial vocabulary table by using a conditional random field-based sequence annotation model to generate a part-of-speech annotation result; performing concept category annotation on the part-of-speech annotation result based on a domain ontology to generate a concept annotation result; and performing relationship annotation on the concept annotation result through dependency syntax analysis to extract inter-word semantic relationships and generate a power grid planning field vocabulary table.

[0016] based on the power grid planning field vocabulary table, constructing a semantic vector space, mapping the power grid planning historical data into the semantic vector space to generate power grid planning basic semantic features; and according to power grid planning business rules, performing multi-dimensional decomposition on the power grid planning basic semantic features to form a multi-dimensional semantic feature library including time dimension features, space dimension features, and business dimension features.

[0017] establishing a power grid planning field semantic vector model, the power grid planning field semantic vector model being constructed by using a word2vec algorithm, the training corpus of the word2vec algorithm including terms in the power grid planning field vocabulary table and context information thereof; obtaining a word vector mapping matrix based on the word2vec algorithm training, the word vector mapping matrix recording a high-dimensional vector representation corresponding to each term; and constructing a semantic vector space by using the word vector mapping matrix;

[0018] performing word segmentation processing on the power grid planning historical data, identifying professional terms based on the power grid planning field vocabulary table, and generating a word segmentation sequence; mapping terms in the word segmentation sequence to the semantic vector space through the word vector mapping matrix to obtain term semantic vectors; and performing weighted combination on the term semantic vectors to generate power grid planning basic semantic features;

[0019] obtaining a power grid planning business rule library, the power grid planning business rule library including a time dimension rule set, a space dimension rule set, and a business dimension rule set; extracting time sequence correlation features from the power grid planning basic semantic features based on the time dimension rule set, the time sequence correlation features describing the evolution law of power grid planning elements over time;

[0020] extracting a spatial distribution feature from the power grid planning basic semantic feature based on the spatial dimension rule set, the spatial distribution feature describing a geographical position relationship and a topological connection relationship of a power grid planning element; extracting a business logic feature from the power grid planning basic semantic feature based on the business dimension rule set, the business logic feature describing a constraint condition and a decision basis in a power grid planning process;

[0021] performing feature fusion on the time sequence correlation feature, the spatial distribution feature and the business logic feature, and establishing a multi-dimension semantic feature index structure; constructing a multi-dimension semantic feature library based on the multi-dimension semantic feature index structure.

[0022] adopting an attention mechanism in the feature fusion layer to perform adaptive weight distribution and feature fusion on features of different dimensions; setting a multi-task learning module in the decision output layer to respectively perform model training for power grid load prediction, grid structure optimization and investment benefit evaluation, including:

[0023] performing adaptive weight distribution on multi-dimension features of the power grid planning in the feature fusion layer, the multi-dimension features including power grid load features, grid structure features and investment benefit features; constructing an attention sub-module for each dimension feature, each attention sub-module including a query vector generation unit, a key vector generation unit and a value vector generation unit;

[0024] inputting the multi-dimension features into the corresponding attention sub-modules respectively, generating query vectors, key vectors and value vectors of the multi-dimension features through the query vector generation unit, the key vector generation unit and the value vector generation unit; performing dot product operation on the query vectors and the key vectors and performing scaling to obtain attention scores; performing Softmax normalization on the attention scores to obtain attention weights of the dimension features;

[0025] performing weighted summation on the attention weights of the dimension features and the value vectors to obtain weighted representations of the dimension features; performing feature fusion on the weighted representations of different dimension features, and calculating correlation strengths between dimensions by using a cross-dimension attention mechanism; performing adaptive weight distribution on different dimension features according to the correlation strengths between the dimensions to generate a fusion feature vector;

[0026] setting a multi-task learning module in the decision output layer, the multi-task learning module including a shared feature extraction layer and a task-specific layer; inputting the fusion feature vector into the shared feature extraction layer to extract high-level semantic features shared between tasks; inputting the high-level semantic features into three parallel task-specific layers respectively to perform model training;

[0027] The three parallel task-specific layers include a load prediction training layer, a grid structure optimization training layer, and an investment benefit evaluation training layer; the load prediction training layer models the power grid load characteristics using a time series deep learning model; the grid structure optimization training layer analyzes the power grid topology structure characteristics using a graph neural network; and the investment benefit evaluation training layer trains the investment decision-making process using a deep reinforcement learning model.

[0028] Receiving real-time power grid planning demand data, inputting the real-time power grid planning demand data into the trained deep learning model; outputting a power grid planning initial scheme based on the deep learning model, the power grid planning initial scheme including load distribution prediction results, grid structure configuration schemes, and investment scale suggestions including:

[0029] Receiving real-time power grid planning demand data through a data receiving module, the real-time power grid planning demand data including power grid operation state data, historical load distribution data, current grid topology data, and investment constraint condition data;

[0030] Pretreating the real-time power grid planning demand data, cleaning the real-time power grid planning demand data to eliminate outliers; aligning the time series of the real-time power grid planning demand data; normalizing the real-time power grid planning demand data to obtain pretreated power grid planning demand data;

[0031] Inputting the pretreated power grid planning demand data into the trained deep learning model, the deep learning model including a data encoding unit and a scheme generation unit; the data encoding unit uses a long short-term memory network to process time series data characteristics, a graph convolution network to process network topology characteristics, and a multilayer perceptron to process investment constraint characteristics; the scheme generation unit generates a power grid planning initial scheme based on the output features of the data encoding unit;

[0032] The deep learning model generates the power grid planning initial scheme through a parallel processing mechanism, the power grid planning initial scheme including: generating load distribution prediction results using a probability prediction method, the load distribution prediction results including mean value prediction and confidence interval of future load; generating grid structure configuration schemes using a graph optimization algorithm, the grid structure configuration schemes including substation site selection schemes and line corridor planning schemes; generating investment scale suggestions based on cost-benefit analysis, the investment scale suggestions including construction cost budget and phased investment suggestions.

[0033] Iteratively optimizing the power grid planning initial scheme using a reinforcement learning algorithm, constructing an optimization objective function according to power grid reliability indicators, economic indicators, and environmental protection indicators, and outputting an optimal power grid planning scheme including:

[0034] The initial scheme of the power grid planning is subjected to data standardization processing to obtain standardized planning data containing substation layout data, line corridor data and investment allocation data; a reinforcement learning optimization model is established based on the standardized planning data, a state space is constructed, and the state space records the current substation operation state, line load state and investment execution state; an action space is constructed, and the action space includes substation capacity adjustment amount, line corridor adjustment amount and investment scale adjustment amount;

[0035] A power grid reliability index is constructed based on the operation data in the state space, and the power supply reliability rate, voltage qualification rate and network loss rate are obtained through power flow calculation and reliability analysis; an economic index is constructed based on the adjustment data in the action space, and the engineering construction cost, operation and maintenance cost and benefit recovery period are obtained through investment benefit analysis; an environmental protection index is constructed based on the combined data of the state space and the action space, and the land occupation area, electromagnetic radiation intensity and carbon emission amount are obtained through environmental impact assessment; the power grid reliability index, economic index and environmental protection index are combined to form an optimization objective function, and the weighted summation method is used to calculate the objective function value, and the objective function value is used as the reward signal of reinforcement learning;

[0036] A deep Q learning network is constructed based on the state space, action space and optimization objective function, and the deep Q learning network maps the current state to the optimal action selection; the state transition sequence after executing the action and the corresponding reward signal are stored in an experience replay pool; training samples are selected from the experience replay pool, and the time difference error of each sample is calculated; the samples are sorted according to the time difference error, and the samples with larger error are preferentially selected for network training;

[0037] The deep Q learning network is updated based on the selected training samples, the loss between the network prediction value and the target value is calculated by using the optimization objective function; the network parameters are optimized through the back propagation algorithm, so that the action selection output by the network is closer to the optimal strategy; the action selection, experience replay and network updating process are repeatedly executed, and when the change of the optimization objective function value in a continuous preset number of rounds is less than a convergence threshold, an optimal power grid planning scheme is output.

[0038] The second aspect of the embodiment of the application,

[0039] A power grid planning construction and operation integrated modeling system based on a semantic system is provided, comprising:

[0040] The first unit is configured to acquire power grid planning historical data, including load prediction data, power grid topology data and equipment parameter data; perform semantic labeling on the power grid planning historical data, and establish a power grid planning field vocabulary; construct a semantic vector space based on the power grid planning field vocabulary, map the power grid planning historical data into the semantic vector space, and generate power grid planning basic semantic features; according to power grid planning business rules, the power grid planning basic semantic features are multi-dimensionally decomposed to form a multi-dimension semantic feature library including time dimension features, space dimension features and business dimension features;

[0041] The second unit is configured to construct a multi-layer neural network structure, including a feature extraction layer, a feature fusion layer and a decision output layer; feature data in the multi-dimension semantic feature library is input into the feature extraction layer, and local features are extracted through convolution operation; in the feature fusion layer, an attention mechanism is used to adaptively allocate weights and fuse features of different dimensions; in the decision output layer, a multi-task learning module is set, and model training is performed respectively for power grid load prediction, grid structure optimization and investment benefit evaluation;

[0042] The third unit is configured to receive real-time power grid planning demand data, input the real-time power grid planning demand data into the trained deep learning model, output a power grid planning initial scheme based on the deep learning model, the power grid planning initial scheme including load distribution prediction results, grid structure configuration scheme and investment scale suggestion, and iteratively optimize the power grid planning initial scheme by using a reinforcement learning algorithm, construct an optimization objective function according to power grid reliability indexes, economic indexes and environmental protection indexes, and output an optimal power grid planning scheme.

[0043] A third aspect of the embodiment of the application,

[0044] An electronic device is provided, comprising:

[0045] a processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0048] A fourth aspect of the embodiment of the application,

[0049] A computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0050] The beneficial effects of the present application are as follows:

[0051] 1. Improve the efficiency of power grid planning: By building a semantic system and multi-dimensional semantic feature library, the historical data of power grid planning is converted into machine-understandable semantic information, and combined with deep learning model for rapid analysis and decision-making, so as to shorten the power grid planning cycle and improve the planning efficiency.

[0052] 2. Improve the scientificity and rationality of power grid planning scheme: Use multi-layer neural network and attention mechanism to fuse multi-dimensional information such as time, space and business, and consider multiple indicators such as power grid reliability, economy and environmental protection, so as to generate more scientific and reasonable power grid planning scheme.

[0053] 3. Realize the integration of power grid planning, construction and operation: Through reinforcement learning algorithm to iteratively optimize the initial scheme, and combined with real-time power grid planning demand data, the generated planning scheme is closer to the actual operation, which promotes the cooperation of planning, construction and operation, and realizes the integrated management. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The flowchart of the power grid planning, construction and operation integrated modeling method based on the semantic system of the embodiment of the present application is shown in

[0055] Figure 2 The structural diagram of the power grid planning, construction and operation integrated modeling system based on the semantic system of the embodiment of the present application is shown in DETAILED DESCRIPTION

[0056] In order to make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be described clearly and completely in combination with the drawings in the embodiment of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] The technical scheme of the present application will be described in detail in the following specific embodiments. The following several specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0058] Figure 1 The flowchart of the power grid planning, construction and operation integrated modeling method based on the semantic system of the embodiment of the present application is shown in Figure 1 The method comprises:

[0059] S11. Obtain power grid planning historical data, the power grid planning historical data comprising load prediction data, power grid topology data, and equipment parameter data; perform semantic labeling on the power grid planning historical data, establish a power grid planning field glossary; construct a semantic vector space based on the power grid planning field glossary, map the power grid planning historical data into the semantic vector space, and generate power grid planning basic semantic features; according to power grid planning business rules, perform multi-dimensional decomposition on the power grid planning basic semantic features, and form a multi-dimensional semantic feature library comprising time dimension features, space dimension features, and business dimension features;

[0060] S12. Construct a multi-layer neural network structure, the multi-layer neural network structure comprising a feature extraction layer, a feature fusion layer, and a decision output layer; input feature data in the multi-dimensional semantic feature library into the feature extraction layer, and extract local features through convolution operation; in the feature fusion layer, adopt an attention mechanism to perform adaptive weight allocation and feature fusion on features of different dimensions; in the decision output layer, set a multi-task learning module, and respectively perform model training for power grid load prediction, grid structure optimization, and investment benefit evaluation;

[0061] S13. Receive real-time power grid planning demand data, input the real-time power grid planning demand data into the trained deep learning model; output a power grid planning initial scheme based on the deep learning model, the power grid planning initial scheme comprising load distribution prediction results, grid structure configuration schemes, and investment scale suggestions; perform iterative optimization on the power grid planning initial scheme using a reinforcement learning algorithm, construct an optimization objective function according to power grid reliability indexes, economic indexes, and environmental protection indexes, and output an optimal power grid planning scheme.

[0062] In an optional implementation, obtaining power grid planning historical data, the power grid planning historical data comprising load prediction data, power grid topology data, and equipment parameter data; performing semantic labeling on the power grid planning historical data, and establishing a power grid planning field glossary comprises:

[0063] Obtaining power grid planning historical data through a data acquisition interface, the power grid planning historical data comprising load prediction data, power grid topology data, and equipment parameter data; establishing a data quality scoring index system for the power grid planning historical data, the data quality scoring index system comprising data integrity indexes, data accuracy indexes, and data timeliness indexes; performing quality evaluation on the power grid planning historical data based on the data quality scoring index system, and generating a quality evaluation report;

[0064] According to the evaluation result in the quality evaluation report, the power grid planning historical data is processed in a hierarchical manner, and the power grid planning historical data is divided into valid data, to-be-repaired data and invalid data; the missing values in the to-be-repaired data are repaired by using an interpolation method based on time sequence correlation, and the abnormal values in the to-be-repaired data are identified and corrected by using statistical characteristics and expert rules to generate repaired data; and the valid data and the repaired data are combined to form preprocessed data.

[0065] A unified data model is designed, the unified data model including a data identification field, an attribute description field and a correlation relationship field; the preprocessed data is mapped to the unified data model through ETL conversion rules to form standardized data; a version identification is established for the standardized data to record data change information and generate a data traceability chain.

[0066] Basic terms of power grid planning are extracted from normative documents, and the basic terms are used as initial entries; a term frequency-inverse document frequency algorithm is used to analyze the standardized data to extract feature words; the feature words are submitted to a field expert for review, and the feature words that pass the review are merged with the initial entries to form an initial word table.

[0067] A multi-level semantic annotation system is constructed, the multi-level semantic annotation system including a part-of-speech annotation layer, a concept category annotation layer and a relationship annotation layer; a conditional random field-based sequence annotation model is used to perform part-of-speech annotation on the initial word table to generate part-of-speech annotation results; the part-of-speech annotation results are annotated based on a domain ontology to generate concept annotation results; the concept annotation results are annotated based on dependency syntax analysis to extract semantic relationships between words and generate a power grid planning field word table.

[0068] Power grid planning historical data is obtained. Data collection interfaces are established with databases such as power grid dispatching systems and planning management systems to obtain historical load prediction data, power grid topology data and equipment parameter data. Load prediction data includes load prediction values at different time granularities, such as hourly, daily or monthly prediction data. Power grid topology data includes the connection relationships and geographic location information of power equipment such as transmission lines, substations and generators. Equipment parameter data includes technical parameters such as rated voltage, rated capacity and impedance of power equipment. For example, historical data from 2020 to 2022 can be exported from the database, including hourly load prediction values, connection information of each substation and parameter data of all transformers.

[0069] Quality assessment of historical data for power grid planning. Establish a data quality scoring index system, including data completeness, data accuracy, and data timeliness. The data completeness index is used to assess whether there are missing values in the data, such as the number and proportion of missing values. The data accuracy index is used to assess the degree of deviation of the data, such as calculating the error between the data and the actual value. The data timeliness index is used to assess the freshness of the data, such as calculating the time delay of the data. For the historical data collected from 2020 to 2022, perform data quality assessment. For example, it is found that the load forecasting data in July 2021 has 5% missing values, some transformer parameter data has abnormal values, and the data timeliness in 2020 is poor. Generate a quality assessment report to record the evaluation results.

[0070] Hierarchical processing of historical data for power grid planning. According to the results of the quality assessment report, the historical data is divided into valid data, repairable data, and invalid data. For example, data with high completeness, accuracy, and timeliness is marked as valid data; data with a small amount of missing values or abnormal values is marked as repairable data; data with too many missing values or obvious errors is marked as invalid data. According to the previous evaluation results, the load forecasting data in July 2021 with a small amount of missing values is marked as repairable data, the transformer data with abnormal parameters is also marked as repairable data, the data in 2020 is marked as invalid data, and the remaining data is marked as valid data.

[0071] Repairing repairable data. For missing values in repairable data, use interpolation methods based on time series correlation to repair. For example, use linear interpolation or spline interpolation to fill in missing values. For abnormal values in repairable data, use statistical characteristics and expert rules to identify and correct. For example, identify abnormal values according to the statistical distribution of the data, and correct them according to expert experience. For example, use linear interpolation to fill in the missing values in the load forecasting data in July 2021, and correct the parameter abnormal transformer data according to the rules established by experts. Merge the repaired data with the valid data to form preprocessed data.

[0072] Establish a unified data model. Design a unified data model to store and manage preprocessed data. The unified data model includes data identification fields, attribute description fields, and association relationship fields. The data identification field is used to uniquely identify each piece of data, such as assigning each data point a unique ID. The attribute description field is used to describe the specific meaning of the data, such as load value, voltage level, device type, etc. The association relationship field is used to describe the association relationship between data, such as the connection relationship between substations and transmission lines.

[0073] The pre-processed data is mapped to a unified data model. The pre-processed data is mapped to a unified data model through ETL (Extract, Transform, Load) conversion rules to form standardized data. For example, historical data in different formats is converted into a unified data format, and the data is stored in a unified database. A version identifier is established for the standardized data, data change information is recorded, and a data traceability chain is generated to track the modification history of the data.

[0074] A power grid planning domain glossary is constructed. Basic terms of power grid planning are extracted from normative documents, industry standards and technical literature as initial terms. For example, terms such as "substation", "transmission line", "load forecasting" are extracted. The TF-IDF (Term Frequency-Inverse Document Frequency) algorithm is used to analyze the standardized data and extract feature words. For example, analyze the words in the standardized data that have a high frequency and a high degree of discrimination, such as "peak-valley difference", "line loss", "transformer capacity", etc. The extracted feature words are submitted to domain experts for review. The feature words that pass the review are merged with the initial terms to form an initial glossary.

[0075] The initial glossary is annotated with multi-level semantics. A multi-level semantic annotation system is constructed, including part-of-speech annotation layer, concept category annotation layer and relationship annotation layer. A conditional random field-based sequence labeling model is used to annotate the initial glossary with part-of-speech, such as "substation" as a noun and "connect" as a verb. Based on the domain ontology, the part-of-speech annotation results are annotated with concept categories, such as "substation" as "power equipment" category. Through dependency syntax analysis, the concept annotation results are annotated with relationships to extract semantic relationships between words, such as the "connection" relationship between "substation" and "transmission line". Finally, a power grid planning domain glossary is generated, which contains information such as the part-of-speech, concept category and semantic relationship of the words.

[0076] The scheme of the present application can:

[0077] Improve data quality: through data quality assessment and repair, the integrity, accuracy and timeliness of the historical data of power grid planning are effectively improved, providing a reliable data foundation for subsequent planning analysis. Improve data utilization efficiency: by establishing a unified data model and standardized data, unified management and sharing of data are realized, avoiding data silos and data redundancy, and improving data utilization efficiency. Promote knowledge accumulation and sharing: by constructing the power grid planning domain glossary, the standardized expression and accumulation of power grid planning domain knowledge are realized, promoting knowledge sharing and inheritance, and providing support for the intelligent development of power grid planning.

[0078] In an alternative embodiment, a semantic vector space is constructed based on the power grid planning field vocabulary, the power grid planning historical data is mapped into the semantic vector space, and power grid planning basic semantic features are generated; the power grid planning basic semantic features are decomposed in multiple dimensions according to power grid planning business rules, and a multi-dimensional semantic feature library including time dimension features, space dimension features and business dimension features is formed, including:

[0079] A power grid planning field semantic vector model is established, which is constructed using a word2vec algorithm, and the training corpus of the word2vec algorithm includes the word entries in the power grid planning field vocabulary and their context information; a word vector mapping matrix is trained based on the word2vec algorithm, which records the high-dimensional vector representation corresponding to each word entry; and a semantic vector space is constructed using the word vector mapping matrix;

[0080] The power grid planning historical data is subjected to word segmentation processing, professional terms are identified based on the power grid planning field vocabulary, and a word segmentation sequence is generated; the word entries in the word segmentation sequence are mapped to the semantic vector space through the word vector mapping matrix to obtain word entry semantic vectors; the word entry semantic vectors are combined with weights to generate power grid planning basic semantic features;

[0081] A power grid planning business rule library is obtained, which includes a time dimension rule set, a space dimension rule set and a business dimension rule set; time sequence correlation features are extracted from the power grid planning basic semantic features based on the time dimension rule set, which describe the evolution law of power grid planning elements over time;

[0082] Space distribution features are extracted from the power grid planning basic semantic features based on the space dimension rule set, which describe the geographical position relationship and topological connection relationship of power grid planning elements; business logic features are extracted from the power grid planning basic semantic features based on the business dimension rule set, which describe the constraint conditions and decision basis in the power grid planning process;

[0083] The time sequence correlation features, the space distribution features and the business logic features are fused to establish a multi-dimensional semantic feature index structure; and a multi-dimensional semantic feature library is constructed based on the multi-dimensional semantic feature index structure.

[0084] The power grid planning semantic feature extraction and multi-dimensional feature library construction method is implemented as follows:

[0085] First, build a vocabulary table for power grid planning. Collect and organize text materials related to power grid planning, such as professional terms, technical specifications, and equipment names. After manual screening and sorting, a vocabulary table containing commonly used terms is formed. For example, the vocabulary table can include terms such as "transformer", "transmission line", "load forecasting", "distribution network planning", etc. This vocabulary table will serve as the basis for subsequent steps.

[0086] Then, build a semantic vector space based on the vocabulary table. Use the word2vec algorithm to train the power grid planning vocabulary table. The training corpus includes terms and their context information in the vocabulary table, which can be extracted from documents, reports, specifications, and other materials related to power grid planning. For example, for the term "transformer", its context information can be "installing a transformer", "transformer capacity", "transformer maintenance", etc. Through the training of the word2vec algorithm, a word vector mapping matrix can be obtained. Using this matrix, a semantic vector space can be constructed, in which each term is represented as a vector in the space, and semantically similar terms are closer in the space.

[0087] Next, process the historical data of power grid planning. Collect and organize the data of power grid planning in previous years, including planning schemes, equipment information, and operation data. Perform word segmentation on these data and identify professional terms based on the previously constructed power grid planning vocabulary table to generate a word segmentation sequence. For example, for a text describing "newly built 110kV transmission line connecting A substation and B substation", after word segmentation and professional term identification, the word segmentation sequence ["newly built", "110kV", "transmission line", "connect", "A substation", "B substation"] can be obtained. Map each term in the word segmentation sequence to the semantic vector space through the word vector mapping matrix to obtain the semantic vector of each term. Then, combine the semantic vectors of these terms with weights to generate the basic semantic features of power grid planning. The way of weighted combination can be adjusted according to the importance, frequency of occurrence, etc. of the terms.

[0088] Obtain the power grid planning business rule library. This rule library contains time dimension rule set, space dimension rule set and business dimension rule set. Time dimension rule set, for example: load growth rate, equipment aging rate, etc.; space dimension rule set, for example: substation distance, line length, etc.; business dimension rule set, for example: power supply reliability requirement, investment budget limit, etc.

[0089] Based on business rules, multi-dimensional features are extracted. Time dimension rule sets are used to extract time sequence correlation features from power grid planning basic semantic features, such as analyzing historical load data to extract load growth trends. Spatial dimension rule sets are used to extract spatial distribution features, such as analyzing the geographical location information of substations and lines to extract power grid topology. Business dimension rule sets are used to extract business logic features, such as analyzing power supply reliability requirements and investment budgets to extract constraint conditions of planning schemes.

[0090] Finally, a multi-dimensional semantic feature library is constructed. The extracted time sequence correlation features, spatial distribution features and business logic features are fused to establish a multi-dimensional semantic feature index structure. For example, features of different dimensions can be combined into a vector, and an index is established to facilitate retrieval. Based on the index structure, a multi-dimensional semantic feature library is constructed to facilitate subsequent power grid planning analysis and decision-making. For example, according to the load growth trend and power grid topology of a certain area, historical planning schemes that meet the conditions can be quickly retrieved to provide reference for new planning schemes.

[0091] The scheme of the present application can:

[0092] Improve power grid planning efficiency: through the construction of semantic vector space and the extraction of multi-dimensional features, complex power grid planning information can be converted into structured data, which is convenient for computer processing and analysis, thereby improving the efficiency of power grid planning. Improve the quality of power grid planning: the multi-dimensional semantic feature library contains rich historical planning information and business rules, which can provide reference for new planning schemes, avoiding the limitations of human experience, thereby improving the quality of power grid planning. Support intelligent power grid planning: the multi-dimensional semantic feature library provides a data basis for intelligent power grid planning, which can support the construction and application of power grid planning models based on artificial intelligence technology, such as using machine learning algorithms to analyze historical planning data, predict future power grid development trends, and automatically generate optimized planning schemes.

[0093] In an optional implementation, an attention mechanism is used in the feature fusion layer to perform adaptive weight allocation and feature fusion on features of different dimensions; a multi-task learning module is set in the decision output layer to perform model training for power grid load forecasting, grid structure optimization and investment benefit evaluation, including:

[0094] In the feature fusion layer, adaptive weight allocation is performed on multi-dimensional features of the power grid planning, including power grid load features, grid structure features and investment benefit features; an attention sub-module is constructed for each dimension feature, and each attention sub-module includes a query vector generation unit, a key vector generation unit and a value vector generation unit;

[0095] The multi-dimensional features are respectively input into the corresponding attention sub-modules, and the query vector generation unit, the key vector generation unit and the value vector generation unit are used to generate the query vector, the key vector and the value vector of the multi-dimensional features; the query vector and the key vector are subjected to dot product operation and scaling to obtain an attention score; the attention score is subjected to Softmax normalization to obtain an attention weight of the dimensional feature;

[0096] The attention weight of the dimensional feature is weighted and summed with the value vector to obtain a weighted representation of the dimensional feature; the weighted representations of different dimensional features are fused, and a cross-dimensional attention mechanism is used to calculate the correlation strength between dimensions; according to the correlation strength between dimensions, adaptive weight distribution is performed on different dimensional features to generate a fusion feature vector;

[0097] A multi-task learning module is arranged in the decision output layer, and the multi-task learning module includes a shared feature extraction layer and a task-specific layer; the fusion feature vector is input into the shared feature extraction layer to extract high-level semantic features shared between tasks; the high-level semantic features are respectively input into three parallel task-specific layers for model training;

[0098] The three parallel task-specific layers include a load prediction training layer, a grid structure optimization training layer and an investment benefit evaluation training layer; the load prediction training layer uses a time series deep learning model to model the power grid load features; the grid structure optimization training layer uses a graph neural network to analyze the power grid topology structure features; the investment benefit evaluation training layer uses a deep reinforcement learning model to train the investment decision-making process.

[0099] The power grid intelligent planning method aims to realize the intelligentization of power grid planning through multi-dimensional feature fusion and multi-task learning of power grid load, grid structure and investment benefit.

[0100] First, collect and preprocess the power grid historical data. For example, collect the power load data, power grid topology structure data and related investment cost and benefit data in the past 5 years. Clean the collected data, remove outliers and missing values, and perform normalization processing, such as scaling the data to 0 to 1, to ensure that different dimensional feature data are comparable.

[0101] Then, multi-dimensional feature representation is constructed. The pre-processed data is converted into a feature vector. For example, the 24-hour load data of each day is taken as a 24-dimensional load feature vector; the adjacency matrix is used to represent the power grid topology structure, and the adjacency matrix is flattened into a vector as a network structure feature vector; the investment cost, expected return and other indicators are combined into a vector as an investment benefit feature vector. Assuming that the load data of a certain day is [100, 110,..., 150], the load feature vector of the day is [100, 110,..., 150]. Assuming that the cost of an investment is 1 million and the expected return is 1.2 million, the investment benefit feature vector is [100, 120].

[0102] Next, feature fusion is performed. An attention sub-module is constructed for the load feature, the network structure feature and the investment benefit feature. Each attention sub-module includes a query vector generation unit, a key vector generation unit and a value vector generation unit. For example, for the load feature vector, the corresponding query vector, key vector and value vector are generated through the three units. The query vector and the key vector are scaled after point multiplication operation to obtain the attention score. The attention score is normalized by Softmax to obtain the attention weight of the load feature. The weighted sum of the attention weight and the value vector is obtained to obtain the weighted representation of the load feature. The same operation is performed on other dimensional features. Then, the cross-dimensional attention mechanism is used to calculate the correlation strength between dimensions, for example, the correlation strength between the load feature and the network structure feature. The weighted sum of the weighted representations of different dimensional features is obtained according to the correlation strength between the dimensions to obtain the fusion feature vector.

[0103] Subsequently, the fusion feature vector is input into a multi-task learning module. The module includes a shared feature extraction layer and a task-specific layer. The shared feature extraction layer is used to extract high-level semantic features shared by tasks. For example, a multi-layer perceptron is used to extract high-level features. The extracted high-level semantic features are input into three parallel task-specific layers for model training. The three task-specific layers are load prediction training layer, network structure optimization training layer and investment benefit evaluation training layer. The load prediction training layer uses a time series deep learning model, such as a long short-term memory network, to model the power grid load feature. The network structure optimization training layer uses a graph neural network to analyze the power grid topology structure feature. The investment benefit evaluation training layer uses a deep reinforcement learning model to train the investment decision-making process.

[0104] Finally, the power grid planning scheme is output. For example, the load prediction results in the future period, the optimized power grid topology structure and the investment benefit evaluation results are output.

[0105] The scheme of the present application can:

[0106] Improve prediction accuracy: Through multi-dimensional feature fusion and attention mechanism, the correlation between different features can be captured, thereby improving the accuracy of load forecasting, network structure optimization and investment benefit evaluation. Optimize planning scheme: Through multi-task learning, load forecasting, network structure optimization and investment benefit evaluation can be considered simultaneously, thereby obtaining a more comprehensive and optimized power grid planning scheme. Enhance decision efficiency: Through intelligent model training and prediction, manual intervention can be reduced, thereby improving the efficiency of power grid planning.

[0107] In an optional implementation, real-time power grid planning demand data is received, and the real-time power grid planning demand data is input into the trained deep learning model; an initial power grid planning scheme is output based on the deep learning model, the initial power grid planning scheme including load distribution prediction results, network structure configuration schemes, and investment scale suggestions.

[0108] Real-time power grid planning demand data is received by a data receiving module, the real-time power grid planning demand data including power grid operation state data, historical load distribution data, current network topology data, and investment constraint condition data;

[0109] The real-time power grid planning demand data is preprocessed, the real-time power grid planning demand data is cleaned to eliminate outliers, the real-time power grid planning demand data is time series aligned, and the real-time power grid planning demand data is normalized to obtain preprocessed power grid planning demand data;

[0110] The preprocessed power grid planning demand data is input into the trained deep learning model, the deep learning model including a data encoding unit and a scheme generation unit; the data encoding unit uses a long short-term memory network to process time series data features, a graph convolution network to process network topology features, and a multi-layer perception machine to process investment constraint features; the scheme generation unit generates an initial power grid planning scheme based on the output features of the data encoding unit;

[0111] The deep learning model generates the initial power grid planning scheme through a parallel processing mechanism, the initial power grid planning scheme including: generating load distribution prediction results using a probabilistic prediction method, the load distribution prediction results including mean value prediction and confidence interval of future load; generating network structure configuration schemes using a graph optimization algorithm, the network structure configuration schemes including substation site selection schemes and line corridor planning schemes; generating investment scale suggestions based on cost-benefit analysis, the investment scale suggestions including construction cost budget and phased investment suggestions.

[0112] The real-time power grid planning demand data is received and input into the trained deep learning model to generate an initial power grid planning scheme, which includes load distribution prediction results, network structure configuration schemes, and investment scale recommendations.

[0113] First, real-time power grid planning demand data is obtained through the data receiving module. These data cover power grid operation state data (such as current power load, voltage level, line load, etc.), historical load distribution data (such as daily / hourly load data over the past few years, and factors affecting load such as weather, holidays, etc.), current network topology data (such as substation location and capacity, transmission line connection and capacity, etc.), and investment constraint condition data (such as budget limit, environmental regulations, land use restrictions, etc.).

[0114] For example, suppose the data we receive includes hourly load data for the past three years in a certain region, the current grid topology structure (containing 2 substations and 5 transmission lines), and a budget of 100 million yuan for investment in the next five years.

[0115] Next, the received data is preprocessed. This includes data cleaning, such as removing outliers and missing values; time series alignment, such as aligning all time series data to the same time resolution; and normalization, such as scaling all data to the same numerical range.

[0116] For example, during data cleaning, we may find that the load data at certain time points is abnormally high, possibly due to data collection errors or special events. We can replace these outliers by using the mean or median. In terms of time series alignment, we can convert all data to hourly level data. In terms of normalization, we can scale all data to between 0 and 1.

[0117] The preprocessed data will be input into the trained deep learning model. The model consists of a data encoding unit and a scheme generation unit. The data encoding unit uses a long short-term memory network (LSTM) to process time series data features (such as historical load data), a graph convolutional network (GCN) to process network topology features (such as current network topology data), and a multi-layer perceptron (MLP) to process investment constraint features (such as budget limits). The output features of these units will be integrated together.

[0118] For example, the LSTM network learns the temporal patterns in historical load data, such as daily and seasonal fluctuations. The GCN network learns the spatial relationships in the grid topology structure, such as the connections between substations and transmission lines. The MLP network learns the influence of investment constraints, such as the impact of budget limits on network structure planning.

[0119] The scheme generation unit generates an initial power grid planning scheme based on the output features of the data encoding unit. This includes generating load distribution prediction results using a probabilistic prediction method, including mean prediction and confidence interval of future load; generating grid structure configuration scheme using graph optimization algorithm, including substation site selection scheme and line corridor planning scheme; and generating investment scale suggestion based on cost-benefit analysis, including construction cost budget and phased investment suggestion.

[0120] For example, the model may predict that the peak load will grow by 20% in the next five years, with a 95% confidence interval. It may also suggest building a new substation at a specific location and constructing new transmission lines to connect to the existing grid. In addition, it may suggest phasing the investment, such as investing 30 million yuan in the first year, 40 million yuan in the second year, and 30 million yuan in the remaining three years.

[0121] The deep learning model generates the initial power grid planning scheme through a parallel processing mechanism, thereby improving computational efficiency.

[0122] The scheme of the present application can:

[0123] Improve power grid planning efficiency: By automatically generating an initial power grid planning scheme, manual intervention and time cost can be significantly reduced, thereby improving power grid planning efficiency. The deep learning model can quickly process a large amount of historical data and real-time data and generate high-quality planning schemes, thereby speeding up the planning cycle. Optimize power grid planning scheme: The deep learning model can learn complex power grid operation rules and investment constraints and generate a more optimal power grid planning scheme. Compared with traditional methods, this technical solution can more accurately predict future load and optimize grid structure configuration, thereby improving the reliability and economy of the power grid. Enhance the adaptability of power grid planning: This technical solution can adapt to changing power grid operating environment and demand. By receiving real-time data and making dynamic adjustments, this scheme can generate a more adaptive power grid planning scheme, thereby better coping with future uncertainties.

[0124] In an optional embodiment, a reinforcement learning algorithm is used to iteratively optimize the initial power grid planning scheme, an optimization objective function is constructed according to power grid reliability indicators, economic indicators and environmental protection indicators, and an optimal power grid planning scheme is output, including:

[0125] The initial scheme of the power grid planning is subjected to data standardization processing to obtain standardized planning data including substation layout data, line corridor data and investment allocation data; a reinforcement learning optimization model is established based on the standardized planning data to construct a state space which records the current substation operation state, line load state and investment execution state; an action space is constructed, which includes substation capacity adjustment, line corridor adjustment and investment scale adjustment;

[0126] Based on the operation data in the state space, a power grid reliability index is constructed, and through power flow calculation and reliability analysis, the power supply reliability rate, voltage qualification rate and network loss rate are obtained; based on the adjustment data in the action space, an economic index is constructed, and through investment benefit analysis, the engineering construction cost, operation and maintenance cost and benefit recovery period are obtained; based on the combined data of the state space and the action space, an environmental protection index is constructed, and through environmental impact assessment, the land occupation area, electromagnetic radiation intensity and carbon emission are obtained; the power grid reliability index, economic index and environmental protection index are combined to form an optimization objective function, and the weighted summation method is used to calculate the objective function value, which is used as the reward signal of reinforcement learning;

[0127] Based on the state space, action space and optimization objective function, a deep Q-learning network is constructed, which maps the current state to the optimal action selection; the state transition sequence after executing the action and the corresponding reward signal are stored in the experience replay pool; training samples are selected from the experience replay pool, and the time difference error of each sample is calculated; the samples are sorted according to the time difference error, and the samples with larger error are preferentially selected for network training;

[0128] Based on the selected training samples, the deep Q-learning network is updated, and the loss between the network prediction value and the target value is calculated using the optimization objective function; the network parameters are optimized through the back propagation algorithm, so that the action selection output by the network is closer to the optimal strategy; the action selection, experience replay and network updating process are repeated, and when the change of the optimization objective function value in a continuous preset number of rounds is less than a convergence threshold, the optimal power grid planning scheme is output.

[0129] The acquisition of the initial scheme of the power grid planning can be achieved by collecting the topological structure, device parameters, load data and other information of the existing power grid, and combining the future power demand prediction and development planning. For example, the location, capacity of substations, path, length, load capacity of transmission lines, and electricity load data in different time periods in a certain area can be collected. In addition, factors such as new energy access and distributed power development can also be considered to form the initial scheme of the power grid planning.

[0130] Next, the initial scheme of power grid planning is standardized. For example, the capacity of the substation is scaled to 0 to 1, the length of the line is scaled to 0 to 1, and the investment amount is scaled to 0 to 1. The purpose of standardization is to eliminate the influence of different data dimensions and units, and to facilitate the subsequent training and optimization of the model. Assuming that the capacity of a substation is 500 megavolt-ampere and the maximum capacity is 1000 megavolt-ampere, the standardized capacity value is 0.5.

[0131] Based on the standardized planning data, a reinforcement learning optimization model is established. The state space records the current operation state of the power grid, such as the load rate of each substation, the load rate of the line, and the amount of investment. The action space includes adjustments to the power grid planning scheme, such as increasing or decreasing the capacity of the substation, adjusting the line path, increasing or decreasing the investment scale, etc. Taking the capacity adjustment of the substation as an example, the action can be defined as increasing or decreasing the capacity by 10%.

[0132] The reliability index, economic index and environmental index of the power grid are constructed. The reliability index can be obtained by power flow calculation and reliability analysis, such as power supply reliability rate, voltage qualification rate and network loss rate. Assuming that the power supply reliability rate of a certain scheme is 99.5% through power flow calculation, the voltage qualification rate is 99%, and the network loss rate is 2%. The economic index can be obtained by investment benefit analysis, such as engineering construction cost, operation and maintenance cost and benefit recovery period. Assuming that the engineering construction cost of a certain scheme is 100 million yuan, the operation and maintenance cost is 1 million yuan per year, and the benefit recovery period is 10 years. The environmental index can be obtained by environmental impact assessment, such as land occupation area, electromagnetic radiation intensity and carbon emission. Assuming that the land occupation area of a certain scheme is 10 hectares, the electromagnetic radiation intensity meets the national standard, and the carbon emission is 1000 tons.

[0133] These indexes are combined into an optimization objective function, and the weighted sum method is used to calculate the objective function value. The weight can be adjusted according to the actual demand, for example, more attention to reliability, then give the reliability index a higher weight.

[0134] A deep Q-learning network is constructed to map the current state to the optimal action selection. The state transition sequence after executing the action and the corresponding reward signal are stored in the experience replay pool. For example, the initial state, the executed action, the new state and the obtained reward are stored. Training samples are selected from the experience replay pool, and the time difference error of each sample is calculated. According to the time difference error, the samples are sorted by importance, and the samples with larger error are preferentially selected for network training.

[0135] The deep Q-learning network is updated based on the selected training samples, and the loss between the network prediction value and the target value is calculated using an optimization objective function. The network parameters are optimized through a backpropagation algorithm to make the action selection of the network output closer to the optimal strategy. The action selection, experience replay and network update processes are repeatedly performed, and when the optimization objective function value changes by less than a convergence threshold for a predetermined number of consecutive rounds, for example, the objective function value changes by less than 0.01 for 100 consecutive rounds, the optimal power grid planning scheme is output.

[0136] The scheme of the present application can:

[0137] Improve power grid planning efficiency: automatically search for the optimal planning scheme through reinforcement learning algorithm, reduce manual intervention, and shorten the planning period. Improve power grid planning quality: comprehensively consider reliability, economy and environmental protection indicators, and find the optimal balance point through intelligent optimization algorithm to develop more scientific and reasonable planning scheme. Reduce the cost of power grid planning: reduce the cost of engineering construction and operation and maintenance by optimizing the structure of power grid and investment strategy, and improve the investment benefit.

[0138] Figure 2 The structure diagram of the power grid planning construction and operation integrated modeling system based on the semantic system of the embodiment of the present application is shown in Figure 2 As shown, the system comprises:

[0139] A first unit is configured to acquire power grid planning historical data, wherein the power grid planning historical data comprises load prediction data, power grid topology data and equipment parameter data; perform semantic labeling on the power grid planning historical data, establish a power grid planning field vocabulary table, construct a semantic vector space based on the power grid planning field vocabulary table, map the power grid planning historical data into the semantic vector space, and generate power grid planning basic semantic features; according to power grid planning business rules, the power grid planning basic semantic features are multi-dimensionally decomposed to form a multi-dimension semantic feature library comprising time dimension features, space dimension features and business dimension features;

[0140] A second unit is configured to construct a multi-layer neural network structure, wherein the multi-layer neural network structure comprises a feature extraction layer, a feature fusion layer and a decision output layer; feature data in the multi-dimension semantic feature library is input into the feature extraction layer, and local features are extracted through convolution operation; in the feature fusion layer, an attention mechanism is used to adaptively allocate weights and fuse features of different dimensions; in the decision output layer, a multi-task learning module is set to train models for power grid load prediction, grid structure optimization and investment benefit evaluation, respectively;

[0141] The third unit is configured to receive real-time power grid planning demand data, input the real-time power grid planning demand data into the trained deep learning model, output an initial power grid planning scheme based on the deep learning model, and perform iterative optimization on the initial power grid planning scheme by using a reinforcement learning algorithm, and output an optimal power grid planning scheme according to a power grid reliability index, an economy index and an environmental protection index.

[0142] A third aspect of the embodiments of the present application,

[0143] An electronic device is provided, comprising:

[0144] A processor;

[0145] A memory for storing processor-executable instructions;

[0146] The processor is configured to invoke the instructions stored in the memory to perform the method described above.

[0147] A fourth aspect of the embodiments of the present application,

[0148] A computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0149] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0150] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for power grid planning, construction, operation and maintenance integrated modeling based on a semantic system, characterized in that, The method comprises the following steps: acquiring power grid planning historical data, the power grid planning historical data comprising load prediction data, power grid topology data and equipment parameter data; performing semantic labeling on the power grid planning historical data to establish a power grid planning field glossary; constructing a semantic vector space based on the power grid planning field glossary, mapping the power grid planning historical data into the semantic vector space to generate power grid planning basic semantic features; according to power grid planning business rules, performing multi-dimensional decomposition on the power grid planning basic semantic features to form a multi-dimensional semantic feature library comprising time dimension features, space dimension features and business dimension features; constructing a multi-layer neural network structure comprising a feature extraction layer, a feature fusion layer and a decision output layer; inputting feature data in the multi-dimensional semantic feature library into the feature extraction layer to extract local features through convolution operation; in the feature fusion layer, adopting an attention mechanism to perform adaptive weight allocation and feature fusion on features of different dimensions; in the decision output layer, setting a multi-task learning module to respectively perform model training for power grid load prediction, grid structure optimization and investment benefit evaluation; receiving real-time power grid planning demand data and inputting the real-time power grid planning demand data into the trained deep learning model; outputting a power grid planning initial scheme based on the deep learning model, the power grid planning initial scheme comprising load distribution prediction results, grid structure configuration schemes and investment scale suggestions; performing iterative optimization on the power grid planning initial scheme by using a reinforcement learning algorithm, constructing an optimization objective function according to power grid reliability indexes, economic indexes and environmental protection indexes, and outputting an optimal power grid planning scheme.

2. The method of claim 1, wherein, acquiring power grid planning historical data, the power grid planning historical data comprising load prediction data, power grid topology data and equipment parameter data; performing semantic labeling on the power grid planning historical data to establish a power grid planning field glossary comprising: acquiring power grid planning historical data through a data acquisition interface, the power grid planning historical data comprising load prediction data, power grid topology data and equipment parameter data; establishing a data quality scoring index system for the power grid planning historical data, the data quality scoring index system comprising data integrity indexes, data accuracy indexes and data timeliness indexes; performing quality evaluation on the power grid planning historical data based on the data quality scoring index system to generate a quality evaluation report; performing hierarchical processing on the power grid planning historical data according to the evaluation results in the quality evaluation report, dividing the power grid planning historical data into valid data, to-be-repaired data and invalid data; repairing missing values in the to-be-repaired data by using an interpolation method based on time sequence correlation, identifying and correcting abnormal values in the to-be-repaired data by using statistical features and expert rules to generate repaired data; merging the valid data and the repaired data to form preprocessed data; A unified data model is designed, which includes a data identification field, an attribute description field and a correlation relationship field; the pre-processed data is mapped to the unified data model through ETL conversion rules to form standardized data; a version identification is established for the standardized data, data change information is recorded, and a data traceability chain is generated; Basic terms of power grid planning are extracted from normative documents, which are used as initial entries; a term frequency-inverse document frequency algorithm is used to analyze the standardized data to extract feature words; the feature words are submitted to domain experts for review, and the feature words that pass the review are merged with the initial entries to form an initial word table; A multi-level semantic annotation system is constructed, which includes a part-of-speech annotation layer, a concept category annotation layer and a relationship annotation layer; a conditional random field-based sequence annotation model is used to perform part-of-speech annotation on the initial word table to generate part-of-speech annotation results; the part-of-speech annotation results are annotated based on domain ontology to generate concept annotation results; the concept annotation results are annotated through dependency syntax analysis to extract semantic relationships between words, and a power grid planning domain word table is generated.

3. The method of claim 1, wherein, Based on the power grid planning domain word table, a semantic vector space is constructed, the power grid planning historical data is mapped into the semantic vector space to generate power grid planning basic semantic features; according to power grid planning business rules, the power grid planning basic semantic features are decomposed in multiple dimensions to form a multi-dimensional semantic feature library including time dimension features, space dimension features and business dimension features: A power grid planning domain semantic vector model is established, which is constructed using a word2vec algorithm, and the training corpus of the word2vec algorithm includes entries in the power grid planning domain word table and their context information; a word vector mapping matrix is trained based on the word2vec algorithm, which records the high-dimensional vector representation corresponding to each entry; a semantic vector space is constructed using the word vector mapping matrix; The power grid planning historical data is processed for word segmentation, professional terms are identified based on the power grid planning domain word table, and a word segmentation sequence is generated; the entries in the word segmentation sequence are mapped to the semantic vector space through the word vector mapping matrix to obtain entry semantic vectors; the entry semantic vectors are combined with weights to generate power grid planning basic semantic features; A power grid planning business rule library is obtained, which includes a time dimension rule set, a space dimension rule set and a business dimension rule set; time sequence correlation features are extracted from the power grid planning basic semantic features based on the time dimension rule set, which describe the evolution law of power grid planning elements over time; extracting a spatial distribution feature from the power grid planning basic semantic feature based on the spatial dimension rule set, the spatial distribution feature describing a geographical position relationship and a topological connection relationship of a power grid planning element; performing feature fusion on the time sequence correlation feature, the spatial distribution feature and the business logic feature to establish a multi-dimensional semantic feature index structure; and constructing a multi-dimensional semantic feature library based on the multi-dimensional semantic feature index structure.

4. The method of claim 1, wherein, adopting an attention mechanism in the feature fusion layer to perform adaptive weight distribution and feature fusion on features of different dimensions; setting a multi-task learning module in the decision output layer to perform model training respectively for power grid load forecasting, grid structure optimization and investment benefit evaluation, including: performing adaptive weight distribution on multi-dimensional features of the power grid planning in the feature fusion layer, the multi-dimensional features including power grid load features, grid structure features and investment benefit features; constructing an attention sub-module for each dimensional feature, each attention sub-module including a query vector generation unit, a key vector generation unit and a value vector generation unit; inputting the multi-dimensional features into the corresponding attention sub-modules respectively, generating query vectors, key vectors and value vectors of the multi-dimensional features through the query vector generation unit, the key vector generation unit and the value vector generation unit; performing dot product operation on the query vectors and the key vectors and performing scaling to obtain attention scores; performing Softmax normalization on the attention scores to obtain attention weights of the dimensional features; performing weighted summation on the attention weights of the dimensional features and the value vectors to obtain weighted representations of the dimensional features; performing feature fusion on the weighted representations of different dimensional features, and calculating correlation strengths between dimensions by using a cross-dimensional attention mechanism; performing adaptive weight distribution on different dimensional features according to the correlation strengths between the dimensions to generate a fusion feature vector; setting a multi-task learning module in the decision output layer, the multi-task learning module including a shared feature extraction layer and task-specific layers; inputting the fusion feature vector into the shared feature extraction layer to extract high-level semantic features shared between tasks; inputting the high-level semantic features into three parallel task-specific layers respectively for model training; the three parallel task-specific layers including a load forecasting training layer, a grid structure optimization training layer and an investment benefit evaluation training layer; the load forecasting training layer adopts a time sequence deep learning model to model power grid load features; the grid structure optimization training layer adopts a graph neural network to analyze power grid topological structure features; and the investment benefit evaluation training layer adopts a deep reinforcement learning model to train an investment decision-making process.

5. The method of claim 1, wherein, receiving real-time power grid planning demand data, and inputting the real-time power grid planning demand data into the trained deep learning model; Output an initial power grid planning scheme based on the deep learning model, the initial power grid planning scheme including load distribution prediction results, network structure configuration schemes, and investment scale suggestions including: Receive real-time power grid planning demand data through a data receiving module, the real-time power grid planning demand data including power grid operation state data, historical load distribution data, current network topology data, and investment constraint condition data; Preprocess the real-time power grid planning demand data, clean the real-time power grid planning demand data to eliminate outliers, align the real-time power grid planning demand data in time series, and normalize the real-time power grid planning demand data to obtain preprocessed power grid planning demand data; Input the preprocessed power grid planning demand data into a trained deep learning model, the deep learning model including a data encoding unit and a scheme generation unit; the data encoding unit uses a long short-term memory network to process time series data features, a graph convolution network to process network topology features, and a multilayer perceptron to process investment constraint features; the scheme generation unit generates an initial power grid planning scheme based on the output features of the data encoding unit; The deep learning model generates the initial power grid planning scheme through a parallel processing mechanism, the initial power grid planning scheme including: generating load distribution prediction results using a probabilistic prediction method, the load distribution prediction results including mean value prediction and confidence interval of future load; generating network structure configuration schemes using a graph optimization algorithm, the network structure configuration schemes including substation site selection schemes and line corridor planning schemes; generating investment scale suggestions based on cost-benefit analysis, the investment scale suggestions including construction cost budget and phased investment suggestions.

6. The method of claim 1, wherein, Iteratively optimize the initial power grid planning scheme using a reinforcement learning algorithm, construct an optimization objective function according to power grid reliability indicators, economic indicators, and environmental protection indicators, and output an optimal power grid planning scheme including: Standardize the initial power grid planning scheme to obtain standardized planning data including substation layout data, line corridor data, and investment allocation data; establish a reinforcement learning optimization model based on the standardized planning data, construct a state space that records the current substation operation state, line load state, and investment execution state; construct an action space including substation capacity adjustment, line corridor adjustment, and investment scale adjustment; The power grid reliability index is constructed based on the operation data in the state space, and the power supply reliability rate, voltage qualification rate and network loss rate are obtained through power flow calculation and reliability analysis; the economic index is constructed based on the adjustment data in the action space, and the engineering construction cost, operation and maintenance cost and benefit recovery period are obtained through investment benefit analysis; the environmental protection index is constructed based on the combined data of the state space and the action space, and the land occupation area, electromagnetic radiation intensity and carbon emission are obtained through environmental impact assessment; the power grid reliability index, economic index and environmental protection index are combined to form an optimization objective function, and the target function value is calculated by weighted summation, which is used as the reward signal of reinforcement learning; The deep Q-learning network is constructed based on the state space, action space and optimization objective function, which maps the current state to the optimal action selection; the state transition sequence after executing the action and the corresponding reward signal are stored in the experience replay pool; the training samples are selected from the experience replay pool, and the time difference error of each sample is calculated; the samples are sorted according to the time difference error, and the samples with larger error are preferentially selected for network training; The deep Q-learning network is updated based on the selected training samples, and the loss between the network prediction value and the target value is calculated by using the optimization objective function; the network parameters are optimized by the back propagation algorithm, so that the action selection output by the network is closer to the optimal strategy; the action selection, experience replay and network update process are repeated, and when the optimization objective function value changes less than the convergence threshold in continuous preset rounds, the optimal power grid planning scheme is output.

7. A semantic system-based integrated modeling system for power grid planning, construction, operation and maintenance, for implementing the method of any one of the preceding claims 1-6, characterized in that, Comprise: The first unit is used for acquiring power grid planning historical data, and the power grid planning historical data includes load prediction data, power grid topology data and equipment parameter data; The power grid planning historical data is semantically labeled, and a power grid planning field glossary is established; Based on the power grid planning field glossary, a semantic vector space is constructed, the power grid planning historical data is mapped into the semantic vector space, and power grid planning basic semantic features are generated; according to the power grid planning business rules, the power grid planning basic semantic features are multi-dimensionally decomposed to form a multi-dimension semantic feature library including time dimension features, space dimension features and business dimension features; The second unit is used for constructing a multi-layer neural network structure, which includes a feature extraction layer, a feature fusion layer and a decision output layer; the feature data in the multi-dimension semantic feature library is input into the feature extraction layer, and local features are extracted through convolution operation; in the feature fusion layer, the attention mechanism is used for adaptive weight allocation and feature fusion of features in different dimensions; in the decision output layer, a multi-task learning module is set, and model training is performed for power grid load prediction, network structure optimization and investment benefit evaluation respectively; The third unit is used for receiving real-time power grid planning demand data, and inputting the real-time power grid planning demand data into the trained deep learning model; Output an initial scheme of power grid planning based on the deep learning model, the initial scheme of power grid planning including a load distribution prediction result, a network structure configuration scheme and an investment scale suggestion; An iterative optimization is performed on the initial scheme of power grid planning by using a reinforcement learning algorithm, an optimization objective function is constructed according to a power grid reliability index, an economy index and an environmental protection index, and an optimal power grid planning scheme is output.

8. An electronic device, comprising: Comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 6.

Images